Zeta

The removal test returns an enormous product. Strip Selene and Tachyon still issues and processes more than 25 million cards across seven countries, still runs core banking, lending, fraud and loyalty, and every bank operation continues, because card processing is deterministic ledger and network code. The models are a servicing layer on infrastructure that predates them.

This is the Clearwater Analytics and LeapXpert position, where a large working platform carries a genuine AI addition, and it is graded C for the same reason. It clears the floor that excluded 10X Banking on a specific distinction worth recording: 10X's models did migration, development and reporting around the ledger and never touched running operations, whereas Selene resolves disputes, processes payments and modifies accounts, which are the bank's operations rather than tooling beside them.

Selene is described as instantly processing payments, resolving disputes and performing account modifications from a cardholder's conversational instruction, which is an agent moving money and changing account state rather than answering questions. One real mechanism is published, orchestration that routes requests between human agents and automated co-pilots with intelligent routing and dynamic scaling, which is a handoff design.

What is absent is every constraint on the automated path: no transaction value threshold, no confidence floor at which a request escalates to a person, no approval gate before an account modification commits, no sampling or post hoc review of automated dispute outcomes, and no statement of which actions are reserved to humans. The routing described is optimised for latency and operational cost, which are throughput objectives rather than control objectives.

No accuracy figure, containment rate, intent classification precision, validation evidence or model documentation was located. The claim that guardrails prevent hallucinations is the strongest assertion made about model behaviour and it carries no measurement, no method and no error analysis, which under the standing approach in this index makes it the least checkable form of a safety claim.

Intent classification is the component that matters most because a misread intent on a conversational channel does not produce a wrong answer, it produces a wrong action against a live account, and no misclassification rate is published for a system that processes payments and modifies accounts.

The platform scale is substantial and checkable: more than 25 million cards live across seven countries, roughly 1,700 staff with about 70 percent in research and development across the United States, Europe, the Middle East, Africa and Asia, and a stated 2 billion dollar valuation reached in 2025.

Investors include SoftBank Vision Fund and Mastercard, the second of which is a card network taking a position in a card processor, which is the buyer adjacent capital shape recorded elsewhere in this index. Celent named the company among the likelier partners for United States banks and credit unions modernising issuer processing.

What is missing is what the axis actually asks for on the artificial intelligence: no institution is named as a Selene deployment, no adoption figure, containment rate or handling time is published, and the card count measures the platform rather than the models. Selene is stated as available to United States customers without naming one.

The scoping architecture is described concretely rather than asserted. Client specific knowledge base embedding brings the issuer's own policies, glossaries and procedures into the model context explicitly to bound scope, and guardrails are described as leveraging issuer specific knowledge and workflows so responses come from approved institutional material rather than open ended generation. That is the same grounded and bounded pattern that earns interface-ai and Hebbia credit.

One caution belongs on the record and it cuts against the honest formulations the index has credited elsewhere: the guardrails are said to prevent hallucinations, which is an elimination claim, where Boosted.ai says reduce and OnFinance AI says resistant. Glia made a comparable claim and backed it with a contractual guarantee; no equivalent commitment appears here.

No data protection agreement, retention schedule, subprocessor list or deletion commitment was located. The exposure is at the top of the range for this index by volume alone, since the platform holds transaction level records for more than 25 million cardholders, and Selene is stated to gain superior performance through real time access to customer data and processing capabilities, which places conversational models directly against live account records rather than a cached summary.

Nothing published addresses what an agent may read, whether conversation transcripts are retained alongside the transaction record, or how cardholder data is handled across the seven jurisdictions involved.

Banking grade security is claimed repeatedly and no attestation, certification, trust centre or enumerated framework was located. The standing note recorded with Persona applies: asserting a security standard without naming one is weaker than silence, because it invites a buyer to assume an attestation exists that they cannot verify.

The absence is more conspicuous here than for a young company, because a card processor at this scale necessarily holds payment card industry attestations to operate at all, so the certifications almost certainly exist and are simply not published anywhere a buyer can read them.

Compliance is asserted repeatedly and never specified. The platform is described as enabling compliant banking asset and liability products, built with banking grade security and compliance, and producing regulatory compliant interactions, with no supervisor, statute, rule, network standard or admission process named anywhere.

The gap is specific for this vendor type: a card processor operating in seven countries sits under payment card industry data security requirements and the operating rules of the card networks, one of which is an investor, and a core banking provider sits inside the supervisory expectations its bank customers are examined against. Naming compliance without naming an instrument is describing the market, the same call recorded for Duna and UPTIQ.

Two exposures, neither addressed. The speech recognition variance finding applies directly and this is its seventh recorded instance: Selene uses speech recognition and text to speech in voice channels, accuracy varies materially by accent, dialect, age and speech impairment, and a cardholder the system cannot understand cannot resolve a dispute or reach their own account, with the fallback path undescribed.

Separately the platform delivers personalised financial advice, context aware recommendations and proactive alerts to cardholders, which is product and behaviour steering at the scale of 25 million cards, and nothing addresses how recommendations are constrained, whether they favour issuer revenue, or how suitability is handled. No per demographic accuracy, accessibility analysis or recommendation governance was located.

No guarantee, indemnity or falsifiable commitment on model behaviour was located, and the contrast with Glia is instructive because Glia makes a similar hallucination claim and puts it in the contract. What lifts this off the floor is regime level recourse the vendor did not create: a cardholder whose dispute is wrongly resolved retains the network chargeback and error resolution rights that apply regardless of whether a human or an agent handled the case, and those rights run against the issuer.

That is allocation rather than vendor accountability, and it does not reach the harm specific to this product, which is an account modification or payment executed on a misread conversational instruction, where nothing published describes detection, reversal or notification.

The architecture is described in unusual detail and the parties are not named. Selene is stated to leverage state of the art foundation models alongside speech recognition and text to speech components, which concedes plainly that external model providers sit in the path, but no provider, hosting arrangement or subprocessor is identified.

Nothing states whether cardholder conversation content and the live account data the agent reads pass to an external model provider during a session, which is the fourth party question an examiner asks and which carries unusual weight here because the same platform holds the system of record.

This vendor does not integrate with the core system, it is the core system, which is the strongest position available on this axis and rare in the index. Tachyon brings issuance, processing, lending, core banking, fraud and risk, loyalty and digital banking together as a single vendor stack, and Selene is described as natively integrated with it, deriving performance from real time access to customer data and processing capabilities rather than from connectors.

The supporting architecture is published in the right terms: cloud native, fully API enabled, real time APIs, granular event streams and extensible data models, with customisable workflows letting issuers integrate their own internal systems and processes. An artificial intelligence layer that sits inside the system of record rather than beside it removes the integration and data freshness problems that constrain almost every other conversational vendor here.

The platform is described as cloud native and operates across seven countries, and no hosting provider, region selection, residency commitment or private deployment option was located. The question is material rather than routine at this scale, because card processing and core banking data are subject to localisation requirements in several of the markets a seven country footprint implies, and because the conversational layer requires real time access to that data. Nothing published states where processing occurs or whether an issuer can constrain it.

No pricing, packaging or basis of charge is published for the platform or for the artificial intelligence layer, and nothing indicates whether Selene is bundled into Tachyon or licensed separately. That second question is the substantive one for a buyer already on the platform, because a conversational layer metered per interaction behaves very differently in cost terms from one included in processing fees, and neither is stated. Category norm for core banking infrastructure rather than a specific failing.

Coverage is unusually complete on three separate dimensions. By institution type the platform serves large banks, credit unions and fintechs, with an independent analyst naming it a likely modernisation partner for United States banks and credit unions specifically.

By product it spans both sides of the balance sheet, cards, loans and deposits, with issuance, processing, lending, core banking, fraud and risk, loyalty and digital banking applications sold as a single vendor stack rather than assembled. By geography it operates across seven countries with delivery presence in the United States, Europe, the Middle East, Africa and Asia. Very few vendors in this index cover asset and liability products for multiple institution types across multiple regions.